AI settings

Overview

This is where what the AI works with lives: access to the providers and the question of which model is used for which task. What the AI knows, by contrast, lives in the CompanyBrain.

Four tabs: Providers · Own & local models · Assignment per task · Agents & prompts. The first three are an administrative matter; the fourth belongs to each individual user and is therefore always there. With the AI team, the tab Web access is added.

One property matters more than any setting: models are not hard-coded. They come from a catalogue queried at the provider and refreshed regularly. A new model therefore appears by itself — without anybody having to ship a new release.

Core tasks

Set up access. For each provider you store an API key. It is stored encrypted and never shown again. Leaving the field empty keeps the stored key unchanged — the field is not a display field but an overwrite field.

Some providers allow a key from the server environment as a fallback. The list states where a key comes from: Key stored, From environment or No key.

Choose a default model. A default model can be set per provider, and one provider can be marked as the overall default.

Connect your own and local models. Providers without a fixed address require a base URL for your own resource. If it is missing, the list reports Address missing. This tab appears only if you hold the necessary permission.

Assignment per task. The third tab decides which model handles which task. That is the lever for cost and quality: a short summary needs no model that a complex analysis would justify.

Let the model be chosen automatically. Instead of a fixed model, the assignment offers Automatic (capability · context · cost). The system then picks, for every request, among the models of the selected provider: if the request needs tools, only models that support tools qualify; request and answer must fit into the model's context; among the suitable ones the one with the lowest estimated cost wins. Free tiers with usage caps are never chosen, and it never switches to another provider. The only basis is what the provider lists in its model list — context length, price per million tokens, capabilities — and the prices you maintain under Model prices (they apply before the provider's figure); these appear in the selection after the model name. If the provider lists none of it (many do not), its default model applies. The Effective column shows the choice for an ordinary request; long requests or tools may lead to another model. Image tasks do not offer this choice.

Maintain model prices. The dollar symbol in a provider's row opens its Model prices: per model the price in US dollars per million tokens for input and output, optionally separate for tokens from the prompt cache (Cache read, Cache write), and per image for image models. The list below shows each model's effective price and where it comes from: maintained, from provider (some providers list prices in their model list) or unknown. A maintained price applies before the provider's figure — for the cost estimate of every call (and with it the AI budget and the log) and for the Automatic choice. Prices are not built into the program (only the per-image prices of the default image models come maintained from the factory, see Automatic media budget): without a maintained price and without a provider figure, the cost of a call is unknown — it is not estimated and does not count towards the budget. Without cache prices, cache tokens count at the input price. Clearing all fields and saving removes the maintained price. Prices apply system-wide; only someone holding the AI management right for all brands may maintain them.

Creating images and videos (Higgsfield). The provider list contains Higgsfield (images and videos). Under Access you enter key ID and secret from the Higgsfield console — both are stored encrypted together and never shown again; if only one is filled in, nothing is stored. Higgsfield only creates images and videos: it is never the default provider, never the target of an assignment per task and never a substitute in the fallback chain. Instead of a default model the row shows how many image and video models are available; maintain their prices as usual under Model prices (per image, per video second).

The default. With Higgsfield connected, it creates images and videos. Without Higgsfield, images come from the existing AI (OpenAI or Gemini, see the assignment Generate an image from a description); videos and images from a reference then get an explanation instead of an error.

How an image or video is created. Create with AI sits at image and file fields of a post and in the media picker. Describe what should be shown, choose the format — 1:1 (post), 4:5 (feed), 9:16 (story, reel), 16:9 (landscape) —, for video the duration and, if you like, a reference image from the file storage (vary an image or turn it into a video). If a model does not support a format it takes the closest one (4:5 for instance as 3:4). Below the description you see before starting what the run costs — Higgsfield's estimate for exactly these settings, otherwise the maintained price, otherwise unknown — and how much AI budget is left. The run works in the background; progress shows Queued, Creating, Saving. When done, the result is in the brand's file storage right away (folder KI-Medien / KI-Bilder or KI-Medien / KI-Videos) and in the post's field; the file carries the AI-generated mark (the basis of the AI label when publishing to networks). Higgsfield keeps results only for a few days — that is why Octibiz downloads them at once.

Cost and budget. Every run counts towards the brand's AI budget (daily and monthly limit, global and per brand). If the estimate plus the brand's runs still in progress no longer fits the budget, the run is rejected — the notice states the amount and what is left. Only a finished run is booked; failed runs and runs rejected under the content rules cost nothing and state their reason. API: POST /api/v1/ai/media-generations/estimate (estimate), POST /api/v1/ai/media-generations (run), GET /api/v1/ai/media-generations/{ulid} (status). Permission: platform.ai_media.create in the brand.

Size of AI videos. Generated videos have their own limit instead of the upload limit — AI images and videos → Largest file per AI video on the Providers tab, 200 MB by default. If a video would be larger by duration and resolution, the run is rejected before it is created; no cost is incurred. Octibiz loads the finished video into the file storage piece by piece and stops as soon as the limit is reached.

Automatic media budget. Images and videos that studios and autopilots create without a click — the visuals of ad variants, the images of posts from the content studio and the weekly-plan autopilot, the images of email versions — run in their own budget per brand and month: AI images and videos → Automatic media budget per brand and month, 50 € by default. It is part of the AI budget — a run has to fit into both. Costs of finished runs and the estimate of running ones count. If the budget is not enough (or set to 0), there is no error: copy and image brief are ready, the visual shows the notice, and the image can be created by hand (with costs shown first). Without a price for the image model, automation creates nothing so the budget holds. So that it works right away without Higgsfield too, the default image models of the existing AI come with a factory price per image — OpenAI gpt-image-1 $0.25 (list price of the highest quality tier, since the model picks the quality itself), Google imagen-3.0-generate-002 $0.03; you can change both under Model prices. Higgsfield states the cost of each run itself. Automation creates video only if the suggestion calls for video — and only via Higgsfield; otherwise it creates an image. GET /api/v1/ai/media-automation/budget returns the state per brand (limit, spent, remaining, rate). The budget counts in euros: providers bill in US dollars, and Octibiz converts estimates and costs at the stored exchange rate of the day (International module, exchange rates). Without a stored rate the dollar amounts count 1:1 — the budget still holds, because a dollar is worth less than a euro; the line below the field says which case applies.

Budget active, but models without a price. If an AI budget is set (daily or monthly limit, global or per brand) and models without a price are in use, this hint appears above the provider list with the models concerned — a click opens the model prices of the provider. Listed are models with calls in the current month whose cost was unknown, and the default model of an active provider access without a price. API: GET /api/v1/ai/budget-status.

Falling back when a provider fails. Each routing can carry a fallback chain; without one of its own the general chain applies. It is entered only when the primary provider fails — server error, overload, timeout. A rejected request (wrong key, disallowed content), an exhausted budget or a blocked address are deliberately not reasons to fall back: a second provider would repeat the same error or bypass the budget limit. If a fallback fails, the next one follows; a fallback without stored access is skipped, while missing access on the primary stays a configuration error. If all fail, the system reports the error of the primary provider — it is the configured one, and the fallback errors are in the log. The budget is checked before every single attempt. Deliberately without a chain are vectors (embeddings) and images: vectors from different models live in incomparable spaces, and a fallback would silently spoil the search index.

Summarise long conversations. Below the routing, the section Long conversations holds the switch Summarise long conversations — on by default. On: when a conversation of the AI assistant gets long, the AI summarises the older messages and sends the summary instead of them (one extra model call roughly every ten messages, its own task Summarise older messages of long conversations). Off: the older messages are only marked as omitted, no extra calls happen, and a summary saved earlier is neither used nor shown.

Maintain agents, skills and prompts. The fourth tab is the library of the AI assistant — the place where both are created, changed, shared and deleted. The assistant links to it from the menu behind the three dots.

Agents are preconfigured assistants: name, instruction, allowed tools, autonomy. Two rules govern what can be set:

• Without an allowed tool an agent is advisory only. Only with tools does the choice between

with confirmation and without asking appear.

• Acting without asking and publishing a writing agent are governance matters. Without the

permission the form does not offer the level, and publishing goes into approval first — the agent stays a draft until then.

Skills. Under Skills you attach skills to an agent. Each selected skill works in two ways:

• Instruction: its text is added after the agent's instruction — in the order of selection, with

the skill's name as heading.

• Tools: the tools the skill requires are added to the agent's allowed tools. An agent without a

tool of its own is then no longer advisory only, and a writing skill counts for approval just like a writing tool.

Only skills (not prompts) that you can see are offered. A colleague's skill may gain tools later; that is why an agent acting without asking writes only with its own tools and those of delivered skills — writing tools from users' skills are always put up for confirmation first. The agent list counts all effective tools under Tools.

Skills and prompts are text building blocks that can be inserted in the assistant's input bar. Some ship with the product; those are read-only because they are restored on every update. For your own version, create a new building block.

Visibility decides in both sections who sees the entry: Private stays with you, Team and Brand share it with others. Entries owned by others appear in the list but cannot be edited.

Web access of the AI team. In the tab Web access (only with the AI team switched on) you choose the search provider (Brave Search, Anthropic web search or Automatic), store the Brave key (encrypted, never shown again), set the daily limits per AI team member, the maximum number of watches per employee (default 20) and the prices per 1,000 searches, and maintain a block list for web pages per brand. What the AI team members do with it is explained in AI team → Web access.

Fields in detail

Every run is recorded with consumption and cost. That is the only way to know at month's end what the money went on.

Settings & permissions

• Every tab hangs on its own permission. Providers and Assignment per task require the AI

management permission, Own & local models the AI workspace governance one. Without the respective permission the tab is not there — not merely locked. Agents & prompts belongs to the user and needs no special permission; the page then opens straight into it. The switch Summarise long conversations can only be changed by users who may also manage system settings (platform.settings.manage).

• Keys cannot be read back. The display shows ••••••••.
• The model catalogue expires and is refreshed. A provider whose catalogue cannot be retrieved

offers no models.

• The assistant additionally needs feature flags. Without ai.gateway and ai.workspace the

AI assistant is unavailable, even with a configured provider.

• Switched off or hidden ⇒ no menu entry. If ai.gateway does not apply to a user, the AI entry

points in the business pages (drafts, summaries, triage, suggestions) and the AI command line in the header are gone for them as well. This settings page stays reachable — setup comes before switching on.

• AI switched off ⇒ a hint instead of silence. While AI is off, users with the AI management

permission see the button AI is off in the header; it leads to this page. At the top it then shows the notice AI is switched off with the button Switch on AI — only users who may also manage system settings (platform.settings.manage) can use it. If no provider access has been entered yet, the notice says so: enter the key under Providers first, then switch on. The AI features appear immediately, without reloading. All other users still see nothing while AI is off.

FAQ & troubleshooting

No models can be selected. Check the provider's access. Without a valid key the catalogue returns nothing.

A new model from the provider is missing. The catalogue is refreshed periodically. Refresh models fetches it immediately.

The assistant says it is not enabled. That is not a question for this page but for the feature flags — see System & feature flags.

The log of an AI call shows costKnown: false. No price is maintained for the model, and the provider lists none. Enter it under Model prices — from then on every call is estimated and counts towards the AI budget.

"Octibiz creates videos through Higgsfield." Higgsfield is not connected. Store key ID and secret under Providers → Higgsfield → Access; until then images without a reference come from the existing AI.

"The AI budget is not enough for this run." The estimate plus the brand's runs still in progress exceeds the remaining budget. Wait for the next budget period or have the budget raised.

A run ends with "Rejected". Higgsfield refused the description or the result under its content rules. Nothing was charged; rephrase the description.

I have forgotten the key. It cannot be read back here. Generate a new one at the provider and enter it.

I cannot edit an agent or building block. Then it is not yours: it was either shared by somebody else or came with the product. Create your own version instead.

A selected skill disappeared after saving. The skill belongs to a different brand than the agent, or you can no longer see it. A private agent without a brand uses only skills without a brand — such as the delivered ones and your private ones.

My agent says "draft" although I chose "published". It uses a writing tool — including one that only comes in through a skill under Skills. Publication is waiting for approval by somebody holding the governance permission.

settings.ai · Available from version 0.5.0